Pauline Frenoy

Les études épidémiologiques qui s'intéressent aux effets des expositions multiples sur la santé humaine doivent relever plusieurs défis méthodologiques, notamment lors de l'analyse statistique. En particulier, les fortes corrélations qui peuvent exister entre les expositions, ainsi que leurs effets potentiellement non-linéaires et non additifs sur la santé entrainent des difficultés de modélisation. L'objectif de cette thèse était d'étudier les effets des expositions aux polluants organiques persistants (POP) sur la santé en explorant différentes méthodes d'analyses statistiques.Deux types de données d'exposition aux POP ont été utilisées chez les femmes de la cohorte française E3N-Générations : des biomarqueurs d'imprégnation sanguine et des estimations d'expositions alimentaires. Différentes méthodes statistiques ont été appliquées afin d'identifier les principaux profils d'imprégnation aux POP, d'étudier les relations entre les niveaux d'imprégnation aux POP et des biomarqueurs de stress oxydatif, et d'explorer les associations entre les expositions alimentaires aux POP et le risque de cancer du sein exprimant le récepteur aux œstrogènes.Six principaux profils d'imprégnation aux POP dans le sang ont été décrits chez 467 femmes de la cohorte E3N. Dans cet échantillon, une méthode bayésienne flexible a permis de modéliser des relations positives, négatives ou nulles, linéaires ou non, entre les niveaux d'imprégnation à 41 POP et les niveaux d'enzymes antioxydantes dans le sang. Dans une échantillon plus large constitué de 66 722 femmes E3N suivies pendant une vingtaine d'années, trois méthodes statistiques fréquentistes ont été utilisées et n'ont révélé aucune association statistiquement significative entre la survenue d'un cancer du sein exprimant les récepteurs aux œstrogènes et 1) les co-expositions alimentaires aux POP fortement corrélés, 2) les profils d'exposition alimentaires aux POP les plus fréquents, et 3) les profils d'exposition alimentaires aux POP les plus liés à l'outcome.Ces travaux ont mis en évidence les défis surmontés par les différentes méthodes utilisées, et leurs limites. Ils illustrent l'importance de choisir les méthodes statistiques les plus appropriées pour répondre aux questions de recherche d'intérêt.

Persistent organic pollutants (POPs) are a group of organic chemical compounds. Contradictory results have emerged in epidemiological studies attempting to elucidate their relationship with breast cancer risk. This study explored the relationship between dietary exposures to multiple POPs and ER-positive breast cancer risk in the French E3N cohort study, using three different approaches to handle multicollinearity among exposures. Intakes of 81 POPs were estimated using food consumption data from a validated semi-quantitative food frequency questionnaire and food contamination data. In the first approach, hierarchical clustering was performed to identify clusters of correlated POPs. For each cluster, the levels of POPs belonging to it were averaged. These average levels were then included in a Cox model to estimate their associations with ER-positive breast cancer occurrence. The second and third approaches applied in the present study were Principal component Cox regression (PCR-Cox) and partial least squares Cox regression (PLS-Cox) respectively, both being dimension-reduction methods (respectively unsupervised and supervised) coupled to a Cox model, used to identify principal components of POPs and to estimate their associations with ER-positive breast occurrence. All models were adjusted for potential confounders previously identified using a directed acyclic graph. The study included 66,722 women with a median follow-up of 20.3 years, during which 3,739 developed an incident ER-positive breast cancer. The variable clustering method did not identify any association between the averaged variables and ER-positive breast cancer risk. Five components were retained using both the PCR-Cox and PLS-Cox methods explaining 82% and 77% of the variance in the initial exposure matrix respectively. Among these components, none was significantly associated with the occurrence of ER-positive breast cancer. This study provides an illustrative example of the application of three distinct statistical methods in the context of highly correlated environmental exposures, discussing their potential relevance and limitations within this specific framework.

Mycotoxins have been hypothesized to contribute to a diversity of adverse health effects in humans, even at low concentrations. Certain mycotoxins are established human carcinogens, whereas for others research suggests potential carcinogenic effects. The aim of this study was to determine the association between dietary exposure to mycotoxins and hepatobiliary cancers in the European Prospective Investigation into Cancer and Nutrition (EPIC) cohort. EPIC questionnaire data were matched to mycotoxin food occurrence data compiled by the European Food Safety Authority to assess long-term dietary mycotoxin exposure (expressed as μg/kg body weight/day) and then relate them to the risk of hepatocellular carcinoma (HCC) (n = 255) and biliary tract cancers (n = 273). Analyses were conducted using multivariable Cox proportional hazards regression models to compute hazard ratios (HR) and 95% confidence intervals (95% CI). Key food groups contributing to mycotoxin exposure were cereals and cereal-based products, vegetables, non-alcoholic beverages (including fruit juices) and fruits. Estimated intake of deoxynivalenol (DON) and its derivatives was positively associated with HCC risk (HR_T3vsT1: 1.90, 95% CI: 1.18–3.05, p-trend <0.01). No statistically significant associations were found for the other mycotoxins. Further research to confirm our observations and investigate potential underlying mechanisms of these compounds is warranted. These data may provide evidence of HCC risks associated with higher dietary intake levels of DON, which has not yet been classified as a human carcinogen.